WAREHOUSE REPLENISHMENT
Loading supplied data…
Click an item to inspect its monthly sales, forecast, floor and order reason. PO CSV includes every item in the selected store/category scope.
DATA ANALYSIS
Search the uploaded sales and stock, inspect order decisions and preview policy changes. Every answer shows its data scope and supporting records.
AI understands your wording; calculations use your uploaded records. Questions use the selected store and category. Product labels, available months and scope go to OpenAI to interpret the question; up to 30 result rows per comparison are sent for the explanation. Uploads stay in this browser session. API credits are required. Guided analysis works without the API. Missing information is explained rather than invented.
Compare the client’s weighted-average method with Croston + MLP using the same stock, course floors, ABC days and order controls.
Largest estimated PO-value changes shown first. Download includes every changed SKU. A lower suggested PO value is not evidence of savings or better service.
Preview a different demand assumption or coverage policy. Your active order settings stay unchanged.
Enter observed before/after results for equal-length periods and the same store/SKU scope. These are separate from forecast estimates.
| Metric | Before AI | During AI pilot |
|---|
Entries stay in this browser session. Download the comparison to retain them. Differences are client-reported observations; they do not prove that AI caused the change.
The August 2026 holdout is one store and one month. Operational savings, expiry loss and service improvements were not measured in that test.
First 100 items shown; the download contains every item. Min and Max retain fractional packs unless the indent rounding control is enabled.
First 100 ranked items shown. Zero-value and negative-average-value items receive C; negative averages are excluded from the ABC share denominator and flagged in the export.
A forecast estimates expected sales in packs. The policy layer converts that estimate into stock limits and an indent. Every method retains ABC day limits and the client’s master course floor.
The Word document repeats the quantity formula under “Indent Val.” This prototype treats it as a wording error and multiplies indent by cost.
Updates average non-zero sale size and the interval between sales separately; forecast = size ÷ interval. α defaults to 0.1. All-zero histories forecast zero. Calculated at monthly frequency because daily data was not supplied.
Classic Croston has bias and does not decay after long runs of zero sales. It does not provide a calibrated service-level confidence interval.
A real pooled neural network with two hidden layers (16 and 8 neurons), trained across the supplied Kalyani SKUs. Inputs: three lags, six-month mean and variability, zero frequency, Croston, calendar month, Core, Chronic and course floor.
“Synth-AI” is a working label, not a supplied vendor or pretrained proprietary system. New uploads use these frozen Kalyani weights; retrain and validate before deploying to other stores.
| Issue | Prototype treatment |
|---|---|
| Two patients vs one patient for low-selling chronic items | Word mode uses the supplied course floor for every SKU. Workbook mode exposes a one/two-patient choice. Zero-sale Non-Core suppression takes priority in workbook mode. |
| Core one-customer stock and Non-Core one-strip stock | Workbook mode uses one pack as an editable design assumption through the master floor; patient-specific quantities need a proper SKU master. |
| On Demand list | 98 supplied codes are tagged. Automatic indent is held by default; disable the hold to see the unrestricted recommendation. |
| Hand-set workbook values and Sun Pharma table | They are historical manual decisions, not inferred rules. The engine recalculates limits; it does not copy those values to future periods. Confirm any manufacturer-specific policy before production. |
| Missing sales rows | Absent SKU-months within the uploaded history are treated as zero observed sales. Confirm export completeness; missing periods and stockouts can conceal demand. |
| Missing master records | Non-Core, not Universal Core, no chronic floor. Shown as missing master. The uploaded working sheet replaces the absent separate Core and Universal Core masters. |
| Pack cost and quantities | Master rate, then arithmetic average batch rate, then zero with a flag. Pack stock is summed across batches; loose units are excluded. |
| Store identification | Branch code joins the data. KALYAN and KALYANI resolve to 86, although filenames mention 85. |
No warehouse availability or open-PO data was supplied. Order value can be understated when cost is missing. This pilot tests calculations; it cannot demonstrate savings across the full network.